RFdiffusion3 is Now Available in foundry

The RFdiffusion3 (RFD3) training and inference code is now available through foundry.

This third-generation model builds on RFdiffusion1 and RFdiffusion2, expanding the scope of what diffusion models can design. RFD3:

• Performs atom-level diffusion across both backbone and side-chain atoms
• Has a new denoising procedure that incorporates a forward step without conditioning information
• Restores symmetry design capabilities introduced in RFdiffusion1
• Introduces new ways to condition your designs

Together, these updates improve the model’s ability to design proteins that interact with small molecules, nucleic acids, and other non-protein partners.

Atom-level diffusion

RFdiffusion3 introduces atoms – not residues – as the fundamental units being diffused. Each residue is modeled with 4 backbone and 10 side chain atoms, with smaller side chains extended using  ‘virtual atoms’ on the Cß. This allows the method to design specific side-chain interactions with ligands and catalytic residues. 

New denoising procedure

Instead of one denoising forward pass, there is now an option in RFD3 to use a weighted average of one pass using conditioning information and one with no conditioning information. This enhances the model’s ability to satisfy complex sets of conditions. 

New and returning types of conditioning

The atomic-resolution in the diffusion procedure allows RFD3 to specify: 

• Hydrogen bond donor/acceptor atoms
• Solvent-accessible surface area labels to define how buried ligand atoms are in the designed proteins
• As in RFdiffusion2, protein designers can specify the center of mass of their designed proteins relative to a target molecule/motif
• Returning from RFdiffusion1, specifying symmetric noise as input will generate symmetric structures

In silico results

• In general, RFD3 generated more diverse structures and docking poses than its precursors 
• The sequences it generates – since RFD3 is side chain aware – are similarly distributed to those generated by MPNN
• RFD3 generates structures in batches, making it ~10× faster than RFD2
• In designing proteins to bind to DNA, RFD3 has a pass rate of 8.67% for monomeric designs and 6.67% for dimeric designs
• For enzyme design, RFD3 outperforms RFD2 on 90% of the atomic motif enzyme benchmark cases

Try it out

We are excited to add this tool to Rosetta Common’s arsenal of protein design methods. You can find the model in foundry. If you have any questions please reach out to the Rosetta Commons development team (Hope Woods, Rachel Clune, Rocco Moretti, Sergey Lyskov) either via Slack or through the Contact Form.